💻 bioinformatics

crispAIPE: Probabilistic Modelling of Prime Editing Variant Correction Efficiency

crispAIPE is a transformer-based probabilistic framework that quantifies uncertainty in prime editing outcomes by modeling competing editing events on a simplex with Dirichlet likelihood and split-conformal calibration, enabling reliable prediction of variant correction efficiency and generalizable design filters across cell types.

Original authors: Ozden, F., Lu, P., Minary, P.

Published 2026-08-03

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Ozden, F., Lu, P., Minary, P.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine you are a master architect trying to build a tiny, invisible bridge inside a living cell. Your goal is to fix a broken blueprint in the cell's instruction manual (its DNA) without accidentally tearing the whole page apart. This is the world of prime editing, a cutting-edge tool that acts like a "search-and-replace" function for our genetic code. Unlike older tools that cut the DNA in half and hope the cell fixes it correctly, prime editing is a precise surgeon that swaps out just the right letters.

However, designing the right "surgical guide" (called a pegRNA) to tell the editor where to go and what to change is incredibly tricky. It's like trying to guess exactly how a specific key will turn in a specific lock without ever having seen the lock before. Sometimes the key works perfectly, sometimes it doesn't turn at all, and sometimes it jams the lock, causing a mess. Scientists have built computer programs to predict how well these keys will work, but until now, those programs only gave a single guess—a "point estimate"—without telling you how much they were sweating over that guess. If a program says a design will work 80% of the time, you have no idea if it's 80% sure or just guessing wildly. In high-stakes science, knowing how uncertain you are is just as important as the prediction itself.

Enter crispAIPE, a new tool developed by researchers at the University of Oxford that changes the game by teaching computers to admit when they aren't sure. Instead of giving a single number, crispAIPE acts like a weather forecaster who doesn't just say "it will rain," but draws a map showing exactly where the rain will fall, how heavy it might be, and how confident they are in that forecast.

The researchers trained this new AI on a massive dataset of over 92,000 different prime-editing experiments. They taught the AI to look at the DNA sequence and the design of the guide RNA, then predict the three possible outcomes: the edit working perfectly, the cell ignoring the edit entirely, or the cell making a messy, unintended mistake. But here is the magic trick: crispAIPE doesn't just guess the most likely outcome; it draws a "confidence bubble" around its prediction. Using a clever statistical method called conformal prediction, the tool guarantees that if it says there is a 95% chance the result will fall inside its bubble, then in real life, the result actually falls inside that bubble 95% of the time. This is a huge leap forward because it gives scientists a safety net. If the bubble is huge and fuzzy, the scientist knows to be careful and maybe test more designs. If the bubble is tiny and tight, they can trust the prediction and move forward with confidence.

The paper shows that crispAIPE is not only good at guessing the right answer (matching the accuracy of the best existing tools) but is also excellent at knowing when it might be wrong. The researchers found that the tool's uncertainty is linked to specific features of the design, such as the chemical makeup of the guide RNA and the size of the DNA change being attempted. By reusing its training on a new set of data from a different cell type, the tool proved it could act as a filter to spot which designs would likely work in new environments. Ultimately, crispAIPE doesn't just predict the future of gene editing; it tells scientists how much they can trust that prediction, turning a game of chance into a more reliable science.

Technical Summary of crispAIPE: Probabilistic Modelling of Prime Editing Variant Correction Efficiency

Problem Statement
Prime editing is a precise genome-editing modality capable of installing substitutions, insertions, and deletions without double-strand breaks or separate donor templates. However, the efficiency of a specific prime editing guide RNA (pegRNA) design is difficult to predict a priori. Existing computational tools (e.g., DeepPE, PRIDICT, DeepPrime) provide only point estimates of editing efficiency. This lack of uncertainty quantification prevents designers from assessing prediction reliability, a critical gap for clinical applications where distinguishing confident predictions from uncertain ones is essential. Furthermore, editing outcomes are compositional: the fractions of edited, unedited, and indel (unintended modification) reads must sum to one, a constraint often ignored by standard regression models.

Methodology
The authors present crispAIPE, a transformer-based probabilistic framework designed to model the joint distribution of prime editing outcomes and quantify prediction uncertainty.

Data and Splitting: The model was trained on 92,423 high-quality pegRNA–target pairs from the PRIDICT Library-1 dataset (HEK293T cells, PE2 editor). To prevent data leakage, the authors employed a mutation-level target-disjoint split, ensuring that all pegRNAs targeting the same mutation were assigned to the same partition (training, validation, or test).
Input Representation: The model utilizes a hybrid encoding strategy. Wild-type and mutated sequences are converted into 5-bit nucleotide tokens. These are combined with binary positional masks for functional regions (protospacer, PBS, RTT-initial, RTT-mutated) and edit directionality channels to form an 11-channel unified representation ( $11 \times L$ , where $L=99$ ).
Architecture: crispAIPE employs a hybrid Transformer-CNN architecture. Nucleotide tokens pass through two stacked transformer encoder layers (with sinusoidal positional encoding) and are concatenated with the unified representation. This combined tensor is processed by a 1D convolutional block and a multilayer perceptron (MLP) to output three parameters.
Probabilistic Modeling: Instead of predicting a single scalar, the model maps inputs to the concentration parameters ( $\alpha_e, \alpha_u, \alpha_i$ ) of a Dirichlet distribution. This naturally models the competitive relationship between the three mutually exclusive outcomes (edited, unedited, indel) on the 2-simplex. The model is trained by minimizing the negative log-likelihood of the Dirichlet distribution.
Uncertainty Quantification (Conformal Prediction): To provide finite-sample coverage guarantees without relying on the Dirichlet model being perfectly calibrated, the authors integrate split-conformal prediction. They construct Highest-Density Regions (HDR) on the 2-simplex. A nonconformity score is calculated for a held-out validation fold to determine a threshold ( $\hat{q}_\gamma$ ). This ensures that the probability of the true outcome falling within the predicted region is at least $1-\gamma$ , regardless of the underlying model's calibration.

Key Results

Predictive Accuracy: On the target-disjoint test set, crispAIPE achieved strong correlations between predicted and observed fractions: Spearman $\rho = 0.835$ (edited), $0.843$ (unedited), and $0.693$ (indel). These results are competitive with state-of-the-art point-prediction tools (PRIDICT, OPED), despite crispAIPE using only sequence and edit context without hand-crafted biophysical features.
Uncertainty Diagnostics: The Dirichlet concentration parameter ( $\alpha_0$ ) and the conformal region area serve as effective uncertainty indicators. Higher $\alpha_0$ correlates with lower prediction error and smaller region areas.
Calibration Performance: The split-conformal HDR construction successfully achieved nominal coverage. At the 95% level, the empirical joint coverage matched the nominal value (95.2%), whereas baseline methods (Bayesian HDR and PCA-ellipse) either under-covered or over-covered. The conformal regions remained valid across different strata of prediction difficulty (indexed by $\alpha_0$ ).
Generalization: When applied to PRIDICT Library-2 (a different cell line context) using the Library-1 calibration quantile unchanged, the conformal region area successfully acted as a filter, identifying pegRNAs whose predictions transferred best.

Significance and Claims
The paper claims that crispAIPE is the first uncertainty-aware framework for prime-editing efficiency prediction. Its primary significance lies in shifting the paradigm from point estimates to full outcome distributions with finite-sample coverage guarantees.

Actionable Reliability: By providing calibrated prediction regions, the tool allows experimentalists to make frequentist statements about design reliability (e.g., "at most 5% of selected designs will fall short of a threshold").
Prioritization Strategy: The framework enables prioritization not just by expected efficiency, but by upper-confidence bounds or by flagging candidates with large uncertainty regions for additional screening before use in sensitive contexts (e.g., stem cells).
Robustness: The method demonstrates that uncertainty quantification can be achieved without requiring the underlying probabilistic model (Dirichlet) to be perfectly calibrated, leveraging the conformal layer to correct for model misspecification.

The authors position this work as a necessary step for high-stakes genome editing applications, where understanding the reliability of a prediction is as critical as the prediction itself.

Technical Summary of crispAIPE: Probabilistic Modelling of Prime Editing Variant Correction Efficiency

More like this